Introduction to Market Combine Mannequin Utilizing Robyn


Whether or not it’s a longtime firm or pretty new available in the market, virtually each enterprise makes use of totally different Advertising and marketing channels like TV, Radio, Emails, Social Media, and so on., to succeed in its potential prospects and improve consciousness about its product, and in flip, maximize gross sales or income.

However with so many advertising channels at their disposal, enterprise must resolve which advertising channels are efficient in comparison with others and extra importantly, how a lot price range must be allotted to every channel. With the emergence of on-line advertising and several other massive information platforms and instruments, advertising is without doubt one of the most outstanding areas of alternatives for information science and machine studying functions.

Studying Goals

  1. What’s Market Combine Modeling, and the way MMM utilizing Robyn is healthier than a standard MMM?
  2. Time Sequence Elements: Pattern, Seasonality, Cyclicity, Noise, and so on.
  3. Promoting Adstocks: Carry-over Impact & Diminishing Returns Impact, and Adstock transformation: Geometric, Weibull CDF & Weibull PDF.
  4. What are gradient-free optimization and Multi-Goal Hyperparameter Optimization with Nevergrad?
  5. Implementation of Market Combine Mannequin utilizing Robyn.

So, with out additional ado, let’s take our first step to know learn how to implement the Market combine mannequin utilizing the Robyn library developed by Fb(now Meta) crew and most significantly, learn how to interpret output outcomes.

This text was revealed as part of the Knowledge Science Blogathon.

Market Combine Modeling (MMM)

It’s to find out the influence of selling efforts on gross sales or market share. MMM goals to establish the contribution of every advertising channel, like TV, Radio, Emails, Social Media, and so on., on gross sales. It helps companies make even handed choices, like on which advertising channel to spend and, extra importantly, what quantity ought to be spent. Reallocate the price range throughout totally different advertising channels to maximise income or gross sales if vital.

What’s Robyn?

It’s an open-source R bundle developed by Fb’s crew. It goals to scale back human bias within the modeling course of by automating vital choices like choosing optimum hyperparameters for Adstocks & Saturation results, capturing Pattern & Seasonality, and even performing mannequin validation. It’s a semi-automated resolution that facilitates the person to generate and retailer totally different fashions within the course of (totally different hyperparameters are chosen in every mannequin), and in flip, supplies us with totally different descriptive and price range allocation charts to assist us make higher choices (not restricted to) about which advertising channels spend on, and extra importantly how a lot ought to be spent on every advertising channel.

How Robyn addresses the challenges of traditional Market Combine Modeling?

The desk beneath outlines how Robyn addresses the challenges of conventional advertising combine modeling.

Earlier than we take deep dive into constructing Market Combine Mannequin utilizing Robyn, let’s cowl some fundamentals that pertain to Market Combine Mannequin.

Time Sequence Elements

You’ll be able to decompose Time sequence information into two elements:

  • Systematic: Elements which have consistency or repetition and will be described and modeled.
  • Non-Systematic: Elements that don’t have consistency, or repetition, and may’t be instantly modeled, for instance, “Noise” in information.
 Fig: Trend Seasonality chart
Fig: Pattern Seasonality chart

Systematic Time Sequence Elements primarily encapsulate the next 3 elements:

  • Pattern
  • Seasonality
  • Cyclicity


When you discover a long-term improve or lower in time sequence information then you’ll be able to safely say that there’s a pattern within the information. Pattern will be linear, nonlinear, or exponential, and it could possibly even change course over time. For e.g., a rise in costs, a rise in air pollution, or a rise within the share value of an organization for a time period, and so on.

 Fig: Plot showing Trend
Fig: Plot exhibiting Pattern

Within the above plot , blue line reveals an upward pattern in information.


When you discover a periodic cycle within the sequence with fastened frequencies then you’ll be able to say there’s a seasonality within the information. These frequencies could possibly be on each day, weekly, month-to-month foundation, and so on. In easy phrases, Seasonality is at all times of a set and identified interval, which means you’ll discover a particular period of time between the peaks and troughs of the info; ergo at occasions, seasonal time sequence is named periodic time sequence too.

For e.g., Retail gross sales going excessive on a couple of explicit festivals or occasions, or climate temperature exhibiting its seasonal habits of being heat days in summer season and chilly days in winter, and so on.

 Fig: Seasonal plot of Air Passengers
Fig: Seasonal plot of Air Passengers

Within the above plot, we are able to discover a powerful seasonality within the months of July and August which means #AirPassengers are highest whereas lowest within the months of Feb & Nov.


While you discover rises and falls, that aren’t of the fastened interval, you’ll be able to say there’s cyclic sample in information. Typically, the common size of cycles can be greater than the size of seasonal patterns. In distinction, the magnitude of cycles tends to be extra inconsistent than that of seasonal patterns.

autoplot(lynx) +xlab("Yr") +ylab("Variety of lynx trapped")
 Fig: #Lynx trapped each year
Fig: #Lynx trapped every year

As we are able to clearly see aperiodic inhabitants cycles of roughly ten years. The cycles will not be of a continuing size – some final 8 or 9 years, and others last more than ten years.


When there’s no Pattern, Cycle, or Seasonality in any way, and if it’s simply random fluctuations in information then we are able to safely say that it’s simply Noise in information.

 Fig: Plot showing Noise
Fig: Plot exhibiting Noise

Within the above plot, there’s no pattern, seasonality, or cyclic habits in any way. They’re very random fluctuations that aren’t predictable and may’t be used to construct a superb Time Sequence Forecasting mannequin.

RoAS(Return on Promoting Spend)

It’s a advertising metric used to evaluate an promoting marketing campaign’s efficacy. ROAS helps companies verify which promoting channels are doing good and the way they will enhance promoting efforts sooner or later to extend gross sales or income. ROAS formulation is:

ROAS= (Income from an advert marketing campaign/ Value of an advert marketing campaign)*100 %

E.g. in the event you spend $2,000 on an advert marketing campaign and also you make $4,000 in revenue, your ROAS can be 200% .

In easy phrases, ROAS represents the income gained from every greenback spent on promoting, and is commonly represented in proportion.

Promoting Adstock

The time period “Adstock “was coined by Simon Broadbent , and it encapsulates two vital ideas:

  • Carryover, or Lagged Impact
  • Diminishing Returns, or Saturation Impact

1. Carryover, or Lagged Impact

Promoting tends to have an impact extending a number of durations after you see it for the primary time. Merely put, an commercial from earlier day, week, and so on. could have an effect on an advert within the present day, week, and so on. It’s referred to as Carryover or lagged Impact.

E.g., Suppose you’re watching a Internet Sequence on YouTube, and a few advert for a product pops up on the display.  You might wait to purchase this product after the business break. It could possibly be as a result of the product is dear, and also you wish to know extra particulars about it, otherwise you wish to examine it with different manufacturers to make a rational choice of shopping for it in the event you want it within the first place. However in the event you see this commercial a couple of extra occasions, it’d have elevated consciousness about this product, and chances are you’ll buy that product. However when you’ve got not seen that advertisements acquire after the primary time, then It’s extremely potential that you just don’t do not forget that sooner or later. That is referred to as the Carryover, or lagged Impact.

You’ll be able to select beneath of the three adstock transformations in Robyn:

  • Geometric
  • Weibull PDF
  • Weibull CDF


This can be a weighted common going again n days, the place n can range by media channel. Essentially the most salient function of the Geometric transformation is its simplicity, contemplating It requires just one parameter referred to as ‘theta’.

For e.g., Let’s say, an promoting spend on day one is $500 and theta = 0.8, then day two has 500*0.7=$400 price of impact carried-over from day one, day three has 400*0.8= $320 from day 2, and so on.

This will make it a lot simpler to speak outcomes to laymen, or non-technical stakeholders. As well as, In comparison with Weibull Distribution(which has two parameters to optimize ), Geometric is far much less computationally costly
& much less time-consuming, and therefore a lot sooner to run.

Robyn’s implementation of Geometric transformation will be written as follows-

 Fig: Robyn's implementation of Geometric transformation
Fig: Robyn’s implementation of Geometric transformation

Weibull Distribution

You do not forget that one particular person geared up with various abilities in your good friend circle, who’ll match into each group. Due to such a dexterous and pliable character, that particular person was a part of virtually each group.

The Weibull distribution is one thing much like that particular person. It might probably match an array of distributions: Regular Distribution, Left-skewed Distribution, and Proper-Skewd Distribution.

You’ll discover 2 variations of a two-parametric Weibull perform: Weibull PDF and Weibull CDF. In comparison with the one-parametric Geometric perform with the fixed “theta”, the Weibull distribution produces time-varying decay charges with the assistance of parameters Form and Scale.

Robyn’s implementation of Weibull distribution will be illustrated conceptually as follows-

 Fig:Robyn's implementation of Weibull distribution
Fig:Robyn’s implementation of Weibull distribution

Weibull’s CDF (Cumulative Distribution Operate)

It has two parameters, form & scale, and has a nonconstant “theta”. Form controls the form of the decay curve, and Scale controls the inflection of the decay curve.

Be aware: The bigger the form, the extra S-shape. The smaller form, the extra L-shape.

Weibull’s PDF (Chance Density Operate)

Additionally has Form & Scale parameters apart from a nonconstant “theta”. Weibull PDF supplies lagged impact.

Fig: Weibull adstock CDF vs PDF

The plot above reveals totally different curves in every plot with totally different values of Form & Scale hyperparameters exhibiting the versatile nature of Weibull adstcoks. Resulting from extra hyperparameters, Weibull adstocks are extra computationally costly than Geometric adstocks. Nonetheless, Weibull PDF is strongly advisable when the product is anticipated to have an extended conversion window.

2. Diminishing Returns Impact/Saturation Impact

Publicity to an commercial creates consciousness concerning the product in shoppers’ thoughts to a sure restrict, however after that influence of commercials to affect shoppers’ buying habits begin diminishing over time. That is referred to as a Saturation impact or Diminishing Returns impact.

Merely put, It’d be presumptuous to say that the extra money you spend on promoting, the upper your gross sales get. In actuality, this development will get weaker the extra we spend.

For instance, growing the YouTube advert spending from $0  to $10,000 will increase our gross sales quite a bit, however growing it from $10,000,000  to $900,000,000  doesn’t do this a lot anymore.


Robyn makes use of the Hill perform to seize the saturation of every media channel.

Hill Operate for Saturation: It’s a two-parametric perform in Robyn . It has two parameters referred to as alpha & gamma. α controls the form of the curve between the exponential and s-shape, and γ (gamma) controls the inflection.

Be aware: bigger the α, the extra S-shape, and the smaller the α, the extra C-shape. Bigger the γ (gamma), the furtherer the inflection within the response curve.

Please take a look at the beneath plots to see how the Hill perform transformation with respect to parameter adjustments:


Ridge Regression

To deal with Multicollinearity in enter information and stop overfitting, Robyn makes use of Ridge Regression to scale back variance. That is aimed toward bettering the predictive efficiency of MMMs.

The mathematical notation for Ridge regression in Robyn is as follows:



Nevergrad is a Python library developed by a crew of Fb. It facilitates the person with derivative-free and evolutionary optimization.


Why Gradient-free Optimization?

It’s simple to compute a perform’s gradient analytically in a couple of circumstances like weight optimization in Neural Networks. Nonetheless, in different circumstances, estimating the gradient will be fairly difficult. For e.g., if perform f is sluggish to compute, non-smooth, time-consuming to judge, or so noisy, strategies that depend on derivates are of little to no use. Algorithms that don’t use derivatives or finite variations are useful in such conditions and are referred to as derivative-free algorithms.

In Advertising and marketing Combine Modeling, we’ve acquired to search out optimum values for a bunch of hyperparameters to search out one of the best mannequin for capturing patterns in our time sequence information.

For e.g., One desires to calculate your media variables’ Adstock and Saturation results. Based mostly in your formulation, one must outline 2 to three hyperparameters per channel. Let’s say we’re modeling 4 totally different media channels plus 2 offline channels. We have now a breakdown of the media channels, making them a complete of 8 channels. So, 8 channels, and a couple of hyperparameters per channel imply you’ll should outline 16 hyperparameters earlier than having the ability to begin the modeling course of.

So, you’ll have a tough time randomly testing all potential mixtures by your self. That’s when Nevergrad says, Maintain my beer.


Nevergrad eases the method of discovering the very best mixture of hyperparameters to reduce the mannequin error or maximize its accuracy.

MOO (Multi-Goal Hyperparameter Optimization) with Nevergrad

Multi-objective hyperparameter optimization utilizing Nevergrad, Meta’s gradient-free optimization platform, is without doubt one of the key improvements in Robyn for implementing MMM. It automates the regularization penalty, adstocking choice, saturation, and coaching dimension for time-series validation. In flip, it supplies us with mannequin candidates with nice predictive energy.

There’re 4 kinds of hyperparameters in Robyn on the time of writing article.

  • Adstocking
  • Saturation
  • Regularization
  • Validation

Robyn Goals to Optimize the three Goal Features:

  • The Normalized Root Imply Sq. Error (NRMSE): also referred to as the Prediction error. Robyn performs time-series validation by spitting the dataset into prepare, validation, and check. nrmse_test is for assessing the out-of-sample predictive efficiency.
  • Decomposition Root Sum of Squared Distance (DECOMP.RSSD): is without doubt one of the key options of Robyn and is aka enterprise error. It reveals the distinction between the share of impact for paid_media_vars (paid media variables), and the share of spend. DECOMP.RSSD can scratch out essentially the most excessive decomposition outcomes. Therefore It helps slim down the mannequin choice.
  • The Imply Absolute Share Error (MAPE.LIFT): Robyn contains another analysis metric referred to as MAPE.LIFT aka Calibration error, once you carry out “mannequin calibration” step. It minimizes the distinction betweenthe causal impact and the expected impact.

Now we perceive the fundamentals of the Market Combine Mannequin and Robyn library. So, let’s begin implementing Market Combine Mannequin (MMM) utilizing Robyn in R.

Step 1: Set up the Proper Packages

#Step 1.a.First Set up required Packages
set up.packages("Robyn")
set up.packages("reticulate")

#Step 1.b Setup digital Surroundings & Set up nevergrad library
py_install("nevergrad", pip = TRUE)
use_virtualenv("r-reticulate", required = TRUE)

If even after set up you’ll be able to’t import Nevergrad then discover your Python file in your system and run beneath line of code by offering path to your Python file.


Now import  the packages and set present working listing.

#Step 1.c Import packages & set CWD


#Step 1.d You'll be able to pressure multi-core utilization by operating beneath line of code
Sys.setenv(R_FUTURE_FORK_ENABLE = "true")
choices(future.fork.allow = TRUE)

# You'll be able to set create_files to FALSE to keep away from the creation of information regionally
create_files <- TRUE

Step 2: Load Knowledge

You’ll be able to load inbuilt simulated dataset or you’ll be able to load your individual dataset.

#Step 2.a Load information

#Step 2.b Load holidays information from Prophet

# Export outcomes to desired listing.
robyn_object<- "~/MyRobyn.RDS"

Step 3: Mannequin Specification

Step 3.1 Outline Enter variables

Since Robyn is a semi-automated software, utilizing a desk just like the one beneath will be useful to assist articulate impartial and Goal variables to your mannequin:

#### Step 3.1: Specify enter variables

InputCollect <- robyn_inputs(
  dt_input = dt_simulated_weekly,
  dt_holidays = dt_prophet_holidays,
  dep_var = "income",
  dep_var_type = "income",
  date_var = "DATE",
  prophet_country = "DE",
  prophet_vars = c("pattern", "season", "vacation"), 
  context_vars = c("competitor_sales_B", "occasions"),
  paid_media_vars = c("tv_S", "ooh_S", "print_S", "facebook_I", "search_clicks_P"),
  paid_media_spends = c("tv_S", "ooh_S", "print_S", "facebook_S", "search_S"),
  organic_vars = "publication", 
  # factor_vars = c("occasions"),
  adstock = "geometric", 
  window_start = "2016-01-01",
  window_end = "2018-12-31",

Signal of coefficients

  • Default: implies that the variable might have both + , or – coefficients relying on the modeling consequence. Nonetheless,
  • Optimistic/Damaging: If you realize the precise influence of an enter variable on Goal variable you then
    can select signal accordingly.

Be aware: All signal management are routinely offered: “+” for natural & media variables and “default” for all others. Nonetheless, you’ll be able to nonetheless customise indicators if vital. 

You can also make use of documentation anytime for extra particulars by operating: ?robyn_inputs

Categorize variables into Natural, Paid Media, and Context variables:

There are 3 kinds of enter variables in Robyn: paid media, natural and context variables. Let’s perceive, learn how to categorize every variable into these three buckets:

  • paid_media_vars
  • organic_vars
  • context_vars

Be aware:

  1. We apply transformation methods to paid_media_vars and organic_vars variables to replicate carryover results and saturation. Nonetheless, context_vars instantly influence the Goal variable and don’t require transformation.
  2. context_vars and organic_vars can settle for both steady or categorical information whereas paid_media_vars can solely settle for steady information.You’ll be able to point out Natural or context variables with categorical information sort underneath factor_vars parameter.
  3. For variables organic_vars and context_vars , steady information will present extra data to the mannequin than categorical. For instance, offering the % low cost of every promotional provide (which is steady information) will present extra correct data to the mannequin in comparison with a dummy variable that reveals the presence of a promotion with simply 0 and 1.

Step 3.2 Specify hyperparameter names and ranges

Robyn’s hyperparameters have 4 elements:

  • Time sequence validation parameter (train_size).
  • Adstock parameters (theta or form/scale).
  • Saturation parameters alpha/gamma).
  • Regularization parameter (lambda).

Specify Hyperparameter Names

You’ll be able to run ?hyper_names to get the precise media hyperparameter names.

hyper_names(adstock = InputCollect$adstock, all_media = InputCollect$all_media)

## Be aware: Set plot = TRUE to supply instance plots for 
#adstock & saturation hyperparameters.

plot_adstock(plot = FALSE)
plot_saturation(plot = FALSE)

# To test most decrease and higher bounds

Specify Hyperparameter Ranges

You’ll have to say higher and decrease bounds for every hyperparameter. For e.g., c(0,0.7). You’ll be able to even point out a scalar worth if you’d like that hyperparameter to be a continuing worth.

# Specify hyperparameters ranges for Geometric adstock
hyperparameters <- listing(
  facebook_S_alphas = c(0.5, 3),
  facebook_S_gammas = c(0.3, 1),
  facebook_S_thetas = c(0, 0.3),
  print_S_alphas = c(0.5, 3),
  print_S_gammas = c(0.3, 1),
  print_S_thetas = c(0.1, 0.4),
  tv_S_alphas = c(0.5, 3),
  tv_S_gammas = c(0.3, 1),
  tv_S_thetas = c(0.3, 0.8),
  search_S_alphas = c(0.5, 3),
  search_S_gammas = c(0.3, 1),
  search_S_thetas = c(0, 0.3),
  ooh_S_alphas = c(0.5, 3),
  ooh_S_gammas = c(0.3, 1),
  ooh_S_thetas = c(0.1, 0.4),
  newsletter_alphas = c(0.5, 3),
  newsletter_gammas = c(0.3, 1),
  newsletter_thetas = c(0.1, 0.4),
  train_size = c(0.5, 0.8)

#Add hyperparameters into robyn_inputs()

InputCollect <- robyn_inputs(InputCollect = InputCollect, hyperparameters = hyperparameters)

Step 3.3 Save InputCollect within the Format of JSON File to Import Later:

You’ll be able to manually save your enter variables and totally different hyperparameter specs in a JSON file which you’ll import simply for additional utilization.

##### Save InputCollect within the format of JSON file to import later
robyn_write(InputCollect, dir = "./")

InputCollect <- robyn_inputs(
  dt_input = dt_simulated_weekly,
  dt_holidays = dt_prophet_holidays,
  json_file = "./RobynModel-inputs.json")

Step 4: Mannequin Calibration/Add Experimental Enter (Non-compulsory)

You need to use Robyn’s Calibration function to extend confidence to pick out your ultimate mannequin particularly once you don’t have details about media effectiveness and efficiency beforehand. Robyn makes use of raise research (check group vs a randomly chosen management group) to know causality of their advertising on gross sales (and different KPIs) and to evaluate the incremental influence of advertisements.

calibration_input <- information.body(
  liftStartDate = as.Date(c("2018-05-01", "2018-04-03", "2018-07-01", "2017-12-01")),
  liftEndDate = as.Date(c("2018-06-10", "2018-06-03", "2018-07-20", "2017-12-31")),
  liftAbs = c(400000, 300000, 700000, 200),
  channel = c("facebook_S",  "tv_S", "facebook_S+search_S", "publication"),
  spend = c(421000, 7100, 350000, 0),
  confidence = c(0.85, 0.8, 0.99, 0.95),
  calibration_scope = c("instant", "instant", "instant", "instant"),
  metric = c("income", "income", "income", "income")

InputCollect <- robyn_inputs(InputCollect = InputCollect, calibration_input = calibration_input)

Step 5: Mannequin Constructing

Step 5.1 Construct Baseline Mannequin

You’ll be able to at all times tweak trials and variety of iterations in line with your enterprise must get one of the best accuracy.You’ll be able to run ?robyn_run to test parameter definition.

#Construct an preliminary mannequin

OutputModels <- robyn_run(
  InputCollect = InputCollect,
  cores = NULL,
  iterations = 2000,
  trials = 5,
  ts_validation = TRUE,
  add_penalty_factor = FALSE

Step 5.2 Mannequin Answer Clustering

Robyn makes use of Ok-Means clustering on every (paid) media variable to search out “greatest fashions” which have NRMSE, DECOM.RSSD, and MAPE(if calibrated was used).


The method for the Ok-means clustering is:

  • When okay = “auto” (which is the default), It calculates the WSS on k-means clustering utilizing okay = 1 to twenty to search out one of the best worth of okay”.
  • After It has run k-means on all Pareto entrance fashions, utilizing the outlined okay, It picks the “greatest fashions” with the bottom normalized mixed errors.

The method for the Ok-means clustering is as follows:

When okay = “auto” (which is the default), It calculates the WSS on k-means clustering utilizing okay = 1 to okay = 20 to search out one of the best worth of okay”.

The method for the Ok-means clustering is:

You’ll be able to run robyn_clusters() to supply listing of outcomes: some visualizations on WSS-k choice, ROI per media on winner fashions.information used to calculate the clusters, and even correlations of Return on Funding (ROI) and so on. . Beneath chart illustrates the clustering choice.

Step 5.3 Prophet Seasonality Decomposition

Robyn makes use of Prophet to enhance the mannequin match and talent to forecast. In case you are undecided about which baselines should be included in modelling, You’ll be able to discuss with the next description:

  • Pattern: Lengthy-term and slowly evolving motion ( growing or lowering course) over time.
  • Seasonality: Seize seasonal behaviour in a short-term cycle, For e.g. yearly.
  • Weekday: Monitor the repeating behaviour on weekly foundation, if each day information is on the market.
  • Vacation/Occasion: Necessary occasions or holidays that extremely influence your Goal variable.

Professional-tip: Customise Holidays & Occasions

Robyn supplies country-specific holidays for 59 international locations from the default “dt_prophet_holidays ” Prophet file already.You need to use dt_holidays parameter to offer the identical data.

In case your nation’s holidays are included otherwise you wish to customise holidays data then you’ll be able to attempt following:

  • Customise vacation dataset: You’ll be able to customise or change the knowledge within the present vacation dataset.You’ll be able to add occasions & holidays into this desk e.g., Black Friday, college holidays, Cyber Monday, and so on.
  • Add a context variable: If you wish to assess the influence of a selected alone then you’ll be able to add that data underneath context_vars variable.

Step 5.4 Mannequin Choice

Robyn leverages MOO of Nevergrad for its mannequin choice step by routinely returning a set of optimum outcomes. Robyn leverages Nevergrad to realize predominant two aims:

  • Mannequin Match: Goals to reduce the mannequin’s prediction error i.e. NRMSE.
  • Enterprise Match: Goals to reduce decomposition distance i.e. decomposition root-sum-square distance (DECOMP.RSSD). This distance metric is for the connection between spend share and a channel’s coefficient decomposition share. If the gap is just too far then its consequence will be too unrealistic -For e.g. promoting channel with the smallest spending getting the most important impact. So this appears type of unrealistic.

You’ll be able to see in beneath chart how Nevergrad rejects most of “dangerous fashions” (bigger prediction error and/or unrealistic media impact). Every blue dot within the chart represents an explored mannequin resolution.


NRMSE on x-axis and DECOMP.RSSD on y-axis are the two features to be minimized. As you’ll be able to discover in beneath chart,with elevated variety of iterations, a pattern down the bottom-left nook is sort of evident.

Based mostly on the NRMSE & DECOMP.RSSD features Robyn will generate a sequence of baseline fashions on the finish of the modeling course of. After reviewing charts and totally different output outcomes, you’ll be able to choose a ultimate mannequin.

Few key parameters that will help you choose the ultimate mannequin:

  • Enterprise Perception Parameters:  You’ll be able to examine a number of enterprise parameters like Return on Funding, media adstock and response curves, share and spend contributions and so on. in opposition to mannequin’s output outcomes. You’ll be able to even examine output outcomes along with your data of trade benchmarks and totally different evluation metrics.
  • Statistical Parameters: If a number of fashions exhibit very related traits in enterprise insights parameters then you’ll be able to choose the mannequin with greatest statistical parameters (e.g. adjusted R-square, and NRMSE being highest and lowest respectively, and so on.).
  • ROAS Convergence Over Iterations Chart: This chart reveals how Return on funding(ROI) for paid media or Return on Advert spend(ROAS) evolves over time and iterations. For few channels, it’s fairly clear that the upper iterations are giving extra “peaky” ROAS distributions, which show larger confidence for sure channel outcomes.

Step 5.5 Export Mannequin Outcomes

## Calculate Pareto fronts, cluster and export outcomes and plots.

OutputCollect <- robyn_outputs(
  InputCollect, OutputModels,
  csv_out = "pareto",
  pareto_fronts = "auto",
  clusters = TRUE,
  export = create_files,
  plot_pareto = create_files,
  plot_folder = robyn_object


You’ll see 4 csv information are exported for additional evaluation:


Interpretation of the Six Charts

1. Response Decomposition Waterfall by Predictor

The chart illustrates the quantity contribution, indicating the proportion of every variable’s impact (intercept + baseline and media variables) on the goal variable. For instance, primarily based on the chart, roughly 10% of the entire gross sales are pushed by the Publication.

Be aware: For established manufacturers/corporations, Intercept and Pattern can account for a good portion of the Response decomposition waterfall chart, indicating that vital gross sales can nonetheless happen with out advertising channel spending.

2. Share of Spend vs. Share of Impact

This chart compares media contributions throughout numerous metrics:

  • Share of spend: Displays the relative spending on every channel.
  • Share of impact: Measures the incremental gross sales pushed by every advertising channel.
  • ROI (Return On Funding): Represents the effectivity of every channel.

When making vital choices, it’s essential to contemplate trade benchmarks and analysis metrics past statistical parameters alone. As an illustration:

  • A channel with low spending however excessive ROI suggests the potential for elevated spending, because it delivers good returns and should not attain saturation quickly as a result of low spending.
  • A channel with excessive spending however low ROI signifies underperformance, however it stays a big driver of efficiency or income. Therefore, spending on this channel ought to be optimized.

Be aware: Decomp.RSSD corresponds to the gap between the share of impact and the share of spend. So, a big worth of Decomp.RSSD could not make reasonable enterprise sense to optimize. Therefore please test this metric whereas evaluating mannequin options.

3. Common Adstock Decay Price Over Time

This chart tells us common % decay fee over time for every channel. Larger decay fee represents the longer impact over time for that particular Advertising and marketing channel.

4. Precise vs. Predicted Response

This plot reveals how effectively the mannequin has predicted the precise Goal variable, given the enter options. We intention for fashions that may seize many of the variance from the precise information, ergo the R-squared ought to be nearer to 1 whereas NRMSE is low.

One ought to try for a excessive R-squared, the place a typical rule of thumb is

  • R squared < 0.8 =mannequin ought to be improved additional;
  • 0.8 < R squared < 0.9 = admissible, however could possibly be improved bit extra;
  • R squared > 0.9 = Good.

Fashions with a low R squared worth will be improved additional by together with a extra complete set of Enter options – that’s, break up up bigger paid media channels or add further baseline (non-media) variables which will clarify the Goal variable(For e.g., Gross sales, Income, and so on.).

Be aware: You’d be cautious of particular durations the place the mannequin is predicting worse/higher. For instance, if one observes that the mannequin reveals noticeably poorer predictions throughout particular durations related to promotional durations, it could possibly function a helpful methodology to determine a contextual variable that ought to be included into the mannequin.

5. Response Curves and Imply Spend by Channel

Response curves for every media channel point out their saturation ranges and may information price range reallocation methods. Channels with sooner curves reaching a horizontal slope are nearer to saturation, suggesting a diminishing return on further spending. Evaluating these curves can assist reallocate spending from saturated to much less saturated channels, bettering general efficiency.

6. Fitted vs. Residual

The scatter plot between residuals and fitted values (predicted values) evaluates whether or not the essential hypotheses/assumptions of linear regression are met, equivalent to checking for homoscedasticity, figuring out non-linear patterns, and detecting outliers within the information.

Step 6: Choose and Save One Mannequin

You’ll be able to examine all exported mannequin one-pagers in final step and choose one which largely displays your enterprise actuality

Step 6: Choose and save anyone mannequin

## Examine all mannequin one-pagers and choose one which largely displays your enterprise actuality.
select_model <- "4_153_2"

ExportedModel <- robyn_write(InputCollect, OutputCollect, select_model, export = create_files)

Step 7: Get Funds Allocation Based mostly on the Chosen Mannequin

Outcomes from price range allocation charts want additional validation. Therefore it’s best to at all times test price range suggestions and talk about along with your shopper.

You’ll be able to apply robyn_allocator() perform to each chosen mannequin to get the optimum price range combine that maximizes the response.

Following are the two situations that you may optimize for:

  • Most historic response: It simulates the optimum price range allocation that may maximize effectiveness or response(Eg. Gross sales, income and so on.) , assuming the identical historic spend;
  • Most response for anticipated spend: This simulates the optimum price range allocation to maximise response or effectiveness, the place you’ll be able to outline how a lot you wish to spend.

For “Most historic response” state of affairs, let’s take into account beneath use case:

Case 1: When each total_budget  & date_range are NULL.

Be aware: It’s default for final month’s spend.

#Get price range allocation primarily based on the chosen mannequin above

# Test media abstract for chosen mannequin

# NOTE: The order of constraints ought to comply with:

AllocatorCollect1 <- robyn_allocator(
  InputCollect = InputCollect,
  OutputCollect = OutputCollect,
  select_model = select_model,
  date_range = NULL,
  state of affairs = "max_historical_response",
  channel_constr_low = 0.7,
  channel_constr_up = c(1.2, 1.5, 1.5, 1.5, 1.5),
  channel_constr_multiplier = 3,
  export = create_files
# Print the price range allocator output abstract

# Plot the price range allocator one-pager

One CSV file can be exported for additional evaluation/utilization.

When you’ve analyzed the mannequin outcomes plots from the listing of greatest fashions, you’ll be able to select one mannequin and cross it distinctive ID to select_model parameter. E.g offering mannequin ID to parameter select_model = “1_92_12” could possibly be an instance of a particular mannequin from the listing of greatest fashions in ‘OutputCollect$allSolutions’ outcomes object.

When you run the price range allocator for the ultimate chosen mannequin, outcomes will probably be plotted and exported underneath the identical folder the place the mannequin plots had been saved.

You’d see plots like below-

Fig: price range allocator chart

Interpretation of the three plots

  1. Preliminary vs. Optimized Funds Allocation: This channel reveals the brand new optimized advisable spend vs authentic spend share. You’ll should proportionally improve or lower the price range for the respective channels of commercial by analysing the distinction between the unique or optimized advisable spend.
  2.  Preliminary vs. Optimized Imply Response: On this chart too, we’ve got optimized and authentic spend, however this time over the entire anticipated response (for e.g., Gross sales). The optimized response is the entire improve in gross sales that you just’re anticipating to have in the event you change budgets following the chart defined above i.e., growing these with a greater share for optimized spend and decreasing spending for these with decrease optimized spend than the unique spend.
  3. Response Curve and Imply Spend by Channel: This chart shows the saturation impact of every channel. It reveals how saturated a channel is ergo, suggests methods for potential price range reallocation. The sooner the curves attain to horizontal/flat slope, or an inflection, the earlier they’ll saturate with every additional $ spent. The triangle denotes the optimized imply spend, whereas the circle represents the unique imply spend.

Step 8: Refresh Mannequin Based mostly on Chosen Mannequin and Saved Outcomes

Following 2 conditions are good match to rebuild the mannequin:

  • Most information is new. As an illustration, If earlier mannequin has 200 weeks of information and 100 weeks new information is added.
  • Add new enter variables or options.
# Present your InputCollect JSON file and ExportedModel specs

json_file <- "E:/DataSciencePrep/MMM/RobynModel-inputs.json"

RobynRefresh <- robyn_refresh(
  json_file = json_file,
  dt_input = dt_simulated_weekly,
  dt_holidays = dt_prophet_holidays,
  refresh_iters = 1500, 
  refresh_trials = 2
  refresh_steps = 14,
# Now refreshing a refreshed mannequin following the identical method
json_file_rf1 <- "E:/DataSciencePrep/MMM/RobynModel-inputs.json"

RobynRefresh <- robyn_refresh(
  json_file = json_file_rf1,
  dt_input = dt_simulated_weekly,
  dt_holidays = dt_prophet_holidays,
  refresh_steps = 8,
  refresh_iters = 1000,
  refresh_trials = 2

# Proceed with new select_model,InputCollect,,and OutputCollect values
InputCollectX <- RobynRefresh$listRefresh1$InputCollect
OutputCollectX <- RobynRefresh$listRefresh1$OutputCollect
select_modelX <- RobynRefresh$listRefresh1$OutputCollect$selectID

Be aware: All the time take into accout to run robyn_write() (manually or routinely) to export present mannequin first for versioning and different utilization earlier than refreshing the mannequin.

Export the 4 CSV outputs within the folder for additional evaluation:



Robyn with its salient options like mannequin calibration & refresh, marginal returns and price range allocation features to supply sooner, extra correct advertising combine modeling (MMM) outputs and enterprise insights does an awesome job. It reduces human bias in modeling course of by automating many of the vital duties.

The three vital takeaways of this text are as follows:

  • With the appearance of Nevergrad, Robyn finds the optimum hyperparameters with out a lot human intervention.
  • With creation of Nevergrad, Robyn finds the optimum hyperparameters with out a lot human intervention.
  • Robyn helps us seize new patterns in information with periodically up to date MMM fashions.


The media proven on this article just isn’t owned by Analytics Vidhya and is used on the Writer’s discretion. 

Related Articles


Please enter your comment!
Please enter your name here

Latest Articles